Research About Head Pose Estimation Based On Manifold Learning

With the rapid development of the cross-discipline of artificial intelligenceand human-computer interaction, computer application is not just as simple asdata calculation, more natural and intelligent computer-human interactionsbecome the future tendency of computer development. With decline inhardware prices, and related technologies become mature, it emerges a varietyof applications on images and video. Image understanding and video analysisare both hot area which are currently being extensively studied, the head poseestimation is the basis of the pose estimation, an important part of the facerecognition, and even an important tool to study human behavior. Therefore,the study about head pose estimation has important research significance andapplication value.Manifold learning method, nonlinear regression method and video trackingmethod are main approaches that have been used to estimate head pose, eachmethod has its own advantages and disadvantages. Manifold learning is a kindof nonlinear dimensionality reduction method, compared with the traditionallinear dimensionality reduction, which can be better able to express theintrinsic relationship between things in the real life. Moreover, it is able tomaintain the high dimensional spatial data distribution characteristics in lowdimension space, get good manifold in low dimension space, especiallysuitable for data visualization. However, manifold learning has a highcomputational complexity, may distort the data distribution of the highdimension space, and even appear singular, unable to get the right results.Nonlinear regression can also be used for nonlinear dimensionality reduction,in the case of model be determined which can get accurate results, but themodeling is a difficult problem.On the basis of research on manifold learning and nonlinear regression usedin the head pose estimation, we analysis the characteristics, advantages anddisadvantages of two methods, and propose the ManiNLR method by combining manifold learning with nonlinear regression. The approachcombine generalized regression neural network, which has the computationalefficiency and high robustness, with some basic manifold learning methods,and an effective nonlinear model is found. The experimental results show thatthis method can effectively improve the head pose estimation accuracy androbustness, and is suitable for various background applications. In the courseof the study, the time consumed in head pose estimation is transferred to themodel training, which is not need to be regularly updated. The new datadiscriminant directly using the linear regression with computational efficiency,the method can reach the effect of real-time. At the same time, on the basis ofthe study about color image, analysis the color information of the face imagesin the head pose estimation, combined with color histograms and significantarea of the image, and applied to the ManiNLR method, can further improvethe accuracy of the method.